Large-scale model-based question answering methods, devices, electronic devices, storage media, agents, and programs

Abstract

To provide a question answering method based on a large-scaled model, an apparatus, an electronic device, a storage medium, an agent and a program, which relate to a technical field of an artificial intelligence, and in particular to technical fields of the large-scaled model, smart search and information processing.SOLUTION: A question answering method includes: inputting, in response to a retrieval content set retrieved based on a question, the question, the retrieval content set and answer generation presentation information into a large-scaled model; and causing the large-scaled model to perform an operation to process a current text corresponding to the retrieval content set based on a current task to be executed in the answer generation presentation information and the question and obtain a processed text, the current task to be executed being determined based on a task execution order in the answer generation presentation information, and an operation to obtain an answer to the question based on the processed text in a case of determining that the processed text meets a predetermined condition.SELECTED DRAWING: Figure 2

Description

【Technical Field】 【0001】 The present disclosure relates to the field of artificial intelligence technology, and in particular, to the technical fields such as large-scale models, intelligent search, information processing, etc. In particular, it relates to a question-answering method, apparatus, electronic device, storage medium, agent and program based on a large-scale model. 【Background Art】 【0002】 With the rapid development of computers and information technology, applications combined with artificial intelligence have already made great progress. For example, natural language processing technology based on large-scale models can understand user questions and provide corresponding answers. However, how to further improve the accuracy of answer generation by large-scale models remains a problem. 【Summary of the Invention】 【Problems to be Solved by the Invention】 【0003】 The present disclosure provides a question-answering method, apparatus, electronic device, storage medium, agent and program based on a large-scale model. 【Means for Solving the Problems】 【0004】 According to one aspect of the present disclosure, a question-answering method based on a large-scale model is provided. The question-answering method based on the large-scale model responds to a search content set retrieved based on a question, and inputs the question, the search content set and answer generation prompt information into the large-scale model. The large-scale model processes the current text corresponding to the search content set based on the current task to be executed in the answer generation prompt information and the question, and obtains the processed text. The current task to be executed is an operation determined based on the task execution order in the answer generation prompt information, and an operation of obtaining an answer to the question based on the processed text when it is determined that the processed text meets a predetermined condition. 【0005】 In other aspects of the present disclosure, a question answering device based on a large model is provided, which inputs the question, the search content set, and answer generation presentation information into a large model in response to a search content set retrieved based on a question, and includes in the large model a processing module that processes the current text corresponding to the search content set based on the current task to be performed in the answer generation presentation information and the question, and obtains the processed text, wherein the current task to be performed is determined based on the task execution order in the answer generation presentation information, and if it is determined that the processed text satisfies predetermined conditions, obtains the answer to the question based on the processed text. 【0006】 According to another aspect of the present disclosure, an electronic device is provided, the electronic device comprising: at least one processor; and a memory communicably connected to the at least one processor, the memory storing instructions executable by the at least one processor, the instructions being executed by the at least one processor so that the at least one processor can perform the method described above. 【0007】 According to other aspects of the present disclosure, a non-temporary computer-readable storage medium is provided which stores computer instructions, and which are used to cause the computer to perform the method described above. 【0008】 According to other aspects of this disclosure, a computer program that, when executed by a processor, accomplishes the above method is provided. 【0009】 According to other aspects of this disclosure, an agent configured to perform the above method is provided. 【0010】 The content described in this section is not intended to identify any key points or important features of the embodiments of this disclosure, nor does it limit the scope of this disclosure. Other features of this disclosure are readily apparent from the following specification. 【0011】 The drawings are for the purpose of better understanding the present invention and do not limit it. [Brief explanation of the drawing] 【0012】 [Figure 1] Figure 1 schematically shows an exemplary system architecture to which the large-scale model-based question answering method and apparatus according to the embodiments of this disclosure can be applied. [Figure 2] Figure 2 schematically shows a flowchart of a question answering method based on a large-scale model according to an embodiment of the present disclosure. [Figure 3] Figure 3 schematically shows a module diagram of a typical question answering system related to a relevant example. [Figure 4] Figure 4 schematically shows a module diagram of a question answering system based on a large-scale model according to an embodiment of the present disclosure. [Figure 5] Figure 5 schematically shows a template diagram of question analysis presentation information according to an embodiment of the present disclosure. [Figure 6] Figure 6 schematically shows a template diagram of answer generation and presentation information according to another embodiment of the present disclosure. [Figure 7] Figure 7 schematically shows a schematic diagram of the effects of an answer according to an embodiment of the present disclosure. [Figure 8] Figure 8 schematically shows an agent according to an embodiment of the present disclosure. [Figure 9] Figure 9 schematically shows a block diagram of a question answering device based on a large-scale model according to an embodiment of the present disclosure. [Figure 10] Figure 10 schematically shows a block diagram of an electronic device according to an embodiment of the present disclosure. [Modes for carrying out the invention] 【0013】 Hereinafter, exemplary embodiments of the present disclosure will be described with reference to the drawings, including various details of the embodiments for the sake of ease of understanding, but these are illustrative only. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for clarity and brevity, descriptions of known functions and structures will be omitted in the following description. 【0014】 Search enhancement systems are a relatively mature technology in current applications and are widely cited in scenarios such as search engine optimization, virtual assistants, smart customer service, and knowledge question answering. By utilizing information from external knowledge libraries, they help large-scale models resolve questions in knowledge acquisition and updating. 【0015】 However, conventional search enhancement systems functionally connect multiple submodules, each of which must perform its own internal data construction, training optimization, and parameter adaptation. Rules and thresholds must be artificially set for the output of each step. This approach lacks consideration of the overall system across submodules, and simultaneously limits the capabilities of the small models associated with each submodule. While it may perform well when executing a single task, when performing multiple related tasks, it can lead to unexpected results. 【0016】 According to the question-answering method based on a large-scale model provided by the present disclosure, in response to a search content set retrieved based on a question, the question, the search content set, and answer generation prompt information are input into the large-scale model. The large-scale model processes the current text corresponding to the search content set based on the current task to be executed in the answer generation prompt information and the question, and obtains the processed text. The current task to be executed is an operation determined based on the task execution order in the answer generation prompt information, and when it is determined that the processed text meets a predetermined condition, an operation of obtaining an answer to the question based on the processed text is executed. By inputting the answer generation prompt information, the question, and the search content set together into the large-scale model, the high processing ability of the large-scale model can be used to collectively execute a plurality of tasks to be executed accumulated in the answer generation prompt information, process the current text, realize end-to-end unified processing, improve the accuracy of answer generation, and reduce the problem of hallucinations occurring in the large-scale model. 【0017】 FIG. 1 schematically shows an exemplary system architecture to which the question-answering method and apparatus based on a large-scale model according to an embodiment of the present disclosure can be applied. 【0018】 Note that FIG. 1 is only an illustration of a system architecture to which embodiments of the present disclosure can be applied for those skilled in the art to understand the technical content of the present disclosure, and does not mean that the embodiments of the present disclosure cannot be applied to other devices, systems, environments, or scenarios. For example, in another embodiment, the exemplary system architecture to which the question-answering method and apparatus based on a large-scale model can be applied may include a terminal device, but the terminal device does not need to interact with the server, and the question-answering method and apparatus based on a large-scale model according to the embodiments of the present disclosure can be realized. 【0019】 As shown in FIG. 1, the system architecture 100 according to this embodiment includes and acquires terminal devices 101, 102, 103, network 104, and server 105. Network 104 provides a medium for communication links between terminal devices 101, 102, 103 and server 105. Network 104 may include various connection types such as, for example, wired and / or wireless communication links. 【0020】 Users can use terminal devices 101, 102, 103 to interact with server 105 via network 104 to send and receive messages and the like. Various communication client applications such as a knowledge browsing system application, a web page browser application, a search system application, an instant messaging tool, a mailbox client, and / or social platform software may be installed on terminal devices 101, 102, 103 (for illustration only). 【0021】 Terminal devices 101, 102, 103 may be various electronic devices having a display and supporting web page browsing, including but not limited to smartphones, tablet computers, laptop computers, and desktop computers. 【0022】 Server 105 may be a server that provides various services, for example, a background management server (for illustration only) that supports the content browsed by users using terminal devices 101, 102, 103. The background management server can perform processing such as analysis on data including questions such as received user requests, and feedback the processing results (for example, web pages, information, answers, or data obtained or generated based on user requests) to the terminal devices. 【0023】 Optionally, server 105 may run one or more agent services or software applications. Users can interact with the agent using terminal devices 101, 102, and 103. 【0024】 The question-answering method according to the embodiments of this disclosure may generally be performed by terminal devices 101, 102, or 103. Accordingly, the question-answering device according to the embodiments of this disclosure may be installed in terminal devices 101, 102, or 103. 【0025】 Alternatively, the question-answering method according to the embodiments of this disclosure may generally be executed by server 105. Accordingly, the question-answering device according to the embodiments of this disclosure may generally be provided on server 105. The question-answering method according to the embodiments of this disclosure may be executed by a server or server cluster that is different from server 105 and can communicate with terminal devices 101, 102, 103 and / or server 105. Accordingly, the question-answering device according to the embodiments of this disclosure may be provided on a server or server cluster that is different from server 105 and can communicate with terminal devices 101, 102, 103 and / or server 105. 【0026】 For example, when a user enters a question via a text box, terminal devices 101, 102, and 103 acquire the entered question, send the acquired question to server 105, where server 105 analyzes the question, determines the search content set, inputs the search content set, the question, and the answer generation and presentation information into a large-scale model, and can acquire the answer to the question. Alternatively, a server or server cluster that can communicate with terminal devices 101, 102, and 103 and / or server 105 can analyze the question and ultimately acquire the answer to the question. The answer to the question can be transmitted to terminal devices 101, 102, and 103 for display to the user, completing the question-answer interaction. 【0027】 It should be understood that the number of terminal devices, networks, and servers in Figure 1 are merely illustrative. Any number of terminal devices, networks, and servers may be used as needed for implementation. 【0028】 In the proposed technology disclosed herein, the collection, storage, use, processing, transmission, provision, publication, and application of relevant user personal information will all comply with the provisions of relevant laws and regulations, employ necessary confidentiality measures, and will not violate public order and morals. 【0029】 In the proposed technologies described herein, user permission or consent was obtained before acquiring or collecting any user personal information. 【0030】 Figure 2 is a flowchart of a question answering method based on a large-scale model according to an embodiment of the present disclosure. 【0031】 As shown in Figure 2, the question answering method may include operations S210 to S230. 【0032】 In operation S210, the question, the search content set, and the answer generation and presentation information are input into the large-scale model in response to the search content set retrieved based on the question. 【0033】 In operation S220, the current text corresponding to the search content set is processed based on the current task and question to be performed in the answer generation presentation information, and the processed text is obtained. 【0034】 In operation S230, if it is determined that the processed text satisfies predetermined conditions, the answer to the question is obtained based on the processed text. 【0035】 According to the embodiments of this disclosure, the question answering method based on a large-scale model may be executed by a server, but is not limited thereto, and may also be executed by a terminal device or an agent. 【0036】 According to embodiments of this disclosure, the search content set may be a collection of relevant information retrieved from multiple data sources based on a question, such as the internet, databases, text sets, social media, etc. It may include, but is not limited to, videos, text, web pages, paragraphs, sentences, or other forms of data, as long as they are relevant to the submitted question. 【0037】 According to embodiments of this disclosure, answer generation suggestion information is used to instruct a large-scale model whether to generate answers based on a provided set of questions and search content. The answer suggestion information can help the large-scale model understand the task requirements it needs to perform and guide it to perform specific operations to generate answers. 【0038】 The elements included in the answer generation and presentation information include, but are not limited to, task roles, task descriptions, task rules, answer generation rules, and output formats. 【0039】 According to some embodiments of this disclosure, a user-provided question can be answered by performing searches across multiple data sources and obtaining a set of searched content. A large-scale model performs tasks to be executed in the answer-providing information, and based on the tasks to be executed, it performs multiple operations on the current text corresponding to the searched content set to obtain processed text that satisfies predetermined conditions. To understand this, if the question is text content, it can be directly input into the large-scale model, and if it is audio or video content, the audio or video content can be converted to corresponding text content before subsequent processing. 【0040】 The current text corresponding to the search content set may include, but is not limited to, the search content set itself, and may also include content that has been processed from the search content set. For example, it may include content obtained by performing a previous task on the search content set using a large-scale model. 【0041】 According to the embodiments of this disclosure, the task to be performed at present is determined based on the task execution order in the answer generation and presentation information. 【0042】 According to embodiments of this disclosure, the task to be performed may be a task that performs a specific operation on the current text, such as generating or extracting text information in a specific format. 【0043】 According to embodiments of this disclosure, the answer generation and presentation information may include multiple tasks to be performed. The task execution order in the answer generation and presentation information determines the order in which these tasks should be executed, thereby ensuring the logic and effectiveness of the entire processing process, and allowing the processed text to be obtained after all tasks to be performed have been completed. 【0044】 According to embodiments of this disclosure, the predetermined conditions may be one or more. They may be set in advance based on the requirements for the answer, or multiple predetermined conditions corresponding to multiple tasks to be processed may be set depending on the tasks to be processed. 【0045】 According to embodiments of this disclosure, if it is determined that the processed text satisfies certain conditions, the processed text can be used as the answer to the question. However, it is not limited to this. The answer may be extracted directly from the processed text, or obtained by performing certain logical inferences, or the answer may be obtained by formatting the processed text. 【0046】 According to the embodiments of this disclosure, by inputting answer generation and presentation information, questions, and search content sets together into a large-scale model, the high processing power of the large-scale model enables the unified execution of multiple tasks to be performed accumulated in the answer generation and presentation information, thereby processing the current text. This achieves end-to-end unified processing, improves the accuracy of answer generation, and reduces questions that cause hallucinations when the large-scale model performs multiple tasks to be performed. 【0047】 The method shown in Figure 2 will be further explained below by combining specific examples with reference to Figures 3 to 7. 【0048】 Figure 3 schematically shows a module diagram of a typical question answering system related to a relevant example. As shown in Figure 3, the question answering system includes an input module 310, an analysis module 320, a search module 330, a generation module 340, and an output module 350. The input module 310 receives questions entered by the user. The analysis module 320 performs search triggers and rewrites the questions. The search system 330 searches the search library, generates summaries, and sorts the searched content. The generation module 340 aggregates the searched content to generate results. The output module 350 outputs the answers. Each module is relatively independent and responsible for a different function, and the inputs and outputs of multiple modules are connected in a pipeline manner. 【0049】 In the above method, each module needs to perform its own internal data construction, training optimization, and parameter adaptation, and the output of each step requires the artificial setting of rules and thresholds, resulting in poor generalizability. Both the summary generation and search sorting models are small-scale models, with weak content completeness and comprehensibility, not only resulting in the loss of key information of search content, but also the fact that low-quality search resources are given priority over high-quality resources in the generative model, resulting in less accurate information that the generative model can refer to. Although the generative module is a large-scale parameter model, the final response quality is poor. If it is necessary to add new consideration features (e.g., temporal features, authoritative features), multiple submodules need to be readjusted, and at the same time, the thresholds of each submodule need to be re-regressed and set, resulting in poor migration and inability to perform a rapid transition. 【0050】 Figure 4 schematically shows a module diagram of a question answering system according to an embodiment of the present disclosure. As shown in Figure 4, the question answering system includes an input module 410, a processing module 420, and an output module 430. The input module 410 receives questions entered by the user, and the processing module 420 processes the entire processing process uniformly. Specifically, it uses a large-scale model to understand the question and determine the search trigger, and when performing a search, it uses the large-scale model to rewrite the question. Then, it performs a search using multiple rewritten questions, inputs the search content set into the large-scale model at once, uses the large-scale model to organize the content set, and generates processed text. The output module 430 outputs the answer. By realizing end-to-end unified processing of the large-scale model, the question answering system as a whole is significantly improved. 【0051】 According to embodiments of the present disclosure, prior to performing operation S210 shown in Figure 2, the question answering method may further include obtaining a search content set. 【0052】 Based on a question, you can retrieve a set of search content from multiple data sources, but this is not limited to that. You can also rewrite a question to retrieve multiple rewritten questions. Based on these multiple rewritten questions, you can retrieve a set of search content from multiple data sources. 【0053】 Selectively, you may run content that rewrites questions using large-scale models. The following describes in detail how to perform question analysis and rewriting tasks using a large-scale model. 【0054】 According to embodiments of this disclosure, in response to a received question, the question and question analysis presentation information are input into a large-scale model, and the large-scale model is instructed to perform the following operations: Based on the search trigger recognition task in the question analysis presentation information, a search trigger analysis is performed on the question, and the search trigger analysis results are obtained. If the search trigger analysis results indicate that a search operation should be triggered, the question is rewritten based on the question rewriting task in the question analysis presentation information, multiple rewritten questions are obtained, and a set of search content is obtained based on the multiple rewritten questions. 【0055】 According to embodiments of this disclosure, the question presentation information may include a set of rules, presentations, or instructions, and may be a predetermined presentation template that forms a search trigger recognition task. A large model can be guided to understand and analyze the question, for example, the large model may determine whether the user's question is timely, whether it is objective knowledge, or whether it exceeds the knowledge boundaries of the large model, and whether the question needs to be searched. 【0056】 For example, a user's question might be, "What's the weather like today?" This question includes a specific time-sensitive requirement, "today," and "What's the weather like?" also exceeds the knowledge boundaries of the large-scale model, requiring the acquisition of the latest weather information from an external weather information platform. In this case, the large-scale model can respond after the search is complete. 【0057】 According to some embodiments of this disclosure, since some user questions may be relatively complex, it is difficult to directly query ideal knowledge in conventional search systems based on semantic matching. Therefore, it is necessary to perform semantic understanding based on the user's original question and rewrite it, thereby obtaining better search results in the search system. The overall principle of question rewriting is that the system first retains complete information and does not lose details that may affect the accuracy of the answer. Furthermore, the question is divided into key parts, core terms or keywords are extracted, and this forms the basis for understanding the question and retrieving relevant answers. 【0058】 For questions with a time-sensitive requirement, it is necessary to rewrite the question to include the corresponding time information. If the original question includes a time-sensitive requirement, first construct a question that includes specific time information to obtain the latest relevant information. Then, rewrite the question without explicit time terms. If the original question does not have clear time terms, construct a question without time terms to broaden the search scope and prevent omission of possible relevant answers. 【0059】 For example, a user's question might be: What were the sales of product A in Japan over the past five months and the past year? This question contains two clear time terms, and rewriting it with a large-scale model yields: "What were the sales of product A in Japan over the past five months?", "What were the sales of product A in Japan over the past five months, July 2024?", "What were the sales of product A in Japan over the past year?", and "What were the sales of product A in Japan over the past year, 2024?". By semantically segmenting the question and expanding the time information, the user's intent can be captured more completely. A rewriting method based on a small-scale model would typically only split the original question into "What were the sales of product A in Japan over the past five months?" and "What were the sales of product A in Japan over the past year?". Such a method lacks expansion for time information, and the search results may be too broad, potentially failing to accurately capture the user's intent. 【0060】 According to the embodiments of this disclosure, by using question analysis presentation information, large-scale models can more effectively process a variety of complex questions. Search trigger analysis and question rewriting increase the diversity and coverage of retrieved information, reduce information gaps, more accurately pinpoint and understand user needs, and provide more accurate answers. 【0061】 According to the embodiments of this disclosure, if the search trigger analysis results indicate that there is no need to trigger a search operation, an answer to the question is generated based on the question. 【0062】 Specifically, if a user's question concerns objective knowledge, does not exceed the knowledge boundaries of the large-scale model itself, and can be directly answered by the knowledge library built into the large-scale model, then external searches are unnecessary, and the large-scale model can provide an accurate, relevant, and useful answer. 【0063】 For example, if a user's question is "What is the diameter of the Earth?", in this case, the answer can be directly provided by the knowledge library built into the large-scale model, without requiring any additional search operations. 【0064】 According to the embodiments of this disclosure, by analyzing the question, it is possible to increase the response speed and provide a faster response time when the response is directly provided by a large-scale model without requiring external searches. 【0065】 According to embodiments of this disclosure, obtaining multiple rewritten questions by rewriting a question based on a question rewriting task in question analysis presentation information includes obtaining multiple rewritten questions by rewriting a question based on a rewriting rule that matches the type of question in the question rewriting task. 【0066】 According to the embodiments of this disclosure, when handling different types of questions, it is necessary to adopt specific, goal-oriented processing principles or policies depending on the particular scene or type to which the question belongs. 【0067】 For example, when rewriting predictive questions, it is necessary to clarify the specific subject of the prediction, the time range of the prediction, the geographical range, etc. The rewritten questions should clearly refer to the future and focus on events, trends, or outcomes that are about to occur or are likely to occur. When rewriting subjective questions, the rewritten questions should focus on an individual's perspective, feeling, or evaluation. The specific subject of the subjective evaluation, such as a product, service, or event, should be clarified. When rewriting in a multilingual context, it is necessary to ensure that the rewritten questions retain the original meaning and context of the original questions. The cultural background and customs of the target language should be taken into consideration, and the rewritten questions should be natural and smooth in the target language. 【0068】 According to the embodiments of this disclosure, by clarifying the mapping relationship between rewriting rules and question types in the question analysis presentation information, it is possible to guide a large-scale model to perform the question rewriting task and rewrite questions using rewriting rules that match the question type, ensuring that the rewritten questions accurately reflect the user's intent and improving the accuracy and efficiency of the search. 【0069】 According to embodiments of this disclosure, obtaining multiple rewritten questions by rewriting a question based on a question rewriting task in question analysis presentation information includes obtaining multiple initial rewritten questions by rewriting a question based on a question rewriting task. If the relevance between each of the multiple initial rewritten questions and the question satisfies the relevance threshold in the question rewriting task, then multiple rewritten questions are obtained based on the multiple initial rewritten questions. 【0070】 According to embodiments of this disclosure, the relationship between each initial rewrite question and the question is an important indicator for evaluating rewrite quality. The semantic similarity between the question and the initial rewrite question can be calculated using natural language processing techniques. By analyzing the keywords of the question and the initial rewrite question, it can be determined whether the initial rewrite question retains key information of the question. 【0071】 According to the embodiments of this disclosure, by setting a relevance threshold for the question analysis presentation information, the large-scale model can independently determine whether the rewritten initial question matches its expectations, select multiple initial rewritten questions, and ensure that the questions used for retrieval accurately reflect the core intent and query needs of the original question, thereby improving the accuracy of the retrieval and the stability of the question responses performed end-to-end by the large-scale model. 【0072】 According to embodiments of this disclosure, metrics for evaluating rewrite quality may further include the availability of rewrites and their impact on the quality of end-to-end generated results. The availability of rewrites represents the effective percentage that can be understood and used for searching by search engines. It can reflect the stability and usability of large-scale model rewrites. End-to-end generated results refer to the entire process from query input to final result output. As part of the overall process, the rewrite quality of queries can directly affect the accuracy and relevance of the final results. Therefore, these two metrics can be used to evaluate the quality of rewrites. 【0073】 Figure 5 schematically shows a template diagram of question analysis presentation information according to an embodiment of the present disclosure. 【0074】 As shown in Figure 5, the question analysis presentation information includes, from top to bottom, a task role 510, a question rewriting task 520, a question rewriting example 530, and a predetermined preset position 540 for the question. The task role 510 is used to define the role that the large-scale model plays. The question rewriting task 520 includes several rewriting principles, also called rewriting rules, including the question completeness assurance principle, the timeliness principle, the predictive rewriting principle, the subjective rewriting principle, and the output principle. Other principles may be deleted or added as needed, and this disclosure is not limited thereto. The question rewriting example 530 is used to clearly display each task example. The pre-set position 540 for the question is located at the bottom, and the user can enter the question at the pre-set position 540 to form complete information that can be used as input data for the large-scale model. 【0075】 The rewriting rules in the question rewriting task clarify the execution criteria for large-scale models and further improve the execution normality of large-scale models. In addition, the example questions or relevance thresholds can clarify whether the rewritten questions output by the large-scale model satisfy the expected effects, serving as evaluation information for when the large-scale model performs the question rewriting task, and as another execution criterion. Based on this, the rewriting rules in the question rewriting task included in the question analysis presentation information, along with evaluation information for assessing whether the rewritten questions satisfy the demand, provide execution criteria to the large-scale model from multiple stages, ensuring end-to-end processing of the large-scale model and reducing questions that cause hallucinations. 【0076】 According to embodiments of this disclosure, inputting questions and question analysis presentation information into a large-scale model includes entering the questions into pre-set positions in the question analysis presentation information and inputting them into the large-scale model. The search trigger recognition task and the question rewriting task in the question analysis presentation information are each given sequence identifiers to indicate the order in which the tasks should be executed. 【0077】 According to the embodiments of this disclosure, in the question analysis presentation information, multiple tasks are arranged in a structured manner, and order identification is added to instruct the execution order of these tasks, thereby enabling the large-scale model to execute the tasks in the correct order. 【0078】 According to the embodiments of this disclosure, a location for entering a question entered by the user can be pre-set in the question analysis presentation information, and the question and the question analysis presentation information can be combined to form a complete input. 【0079】 According to the embodiments of this disclosure, by filling in the question at a predetermined position in the question analysis presentation information to form a complete input, the large-scale model can understand it, and by setting the order identification of the search trigger recognition task and the question rewriting task, the large-scale model can perform the corresponding search trigger recognition and question rewriting operations based on the determined tasks and order, thereby improving the execution efficiency and controllability of the large-scale model in the search trigger recognition and question rewriting operations. 【0080】 The above describes in detail the tasks performed by the large-scale model, such as search trigger recognition and question rewriting. Below, we will describe the answer generation task performed by the large-scale model. 【0081】 According to embodiments of this disclosure, inputting a question, a search content set, and answer generation presentation information into a large-scale model includes inputting the question and the search content set into pre-configured locations within the answer generation presentation information, respectively, and inputting them into the large-scale model. The presentation information includes multiple tasks, each task having an added sequence identifier to indicate the order in which the tasks are executed. 【0082】 According to embodiments of this disclosure, in the answer generation and presentation information, multiple tasks are arranged in a structured manner, and order identification is added to indicate the execution order of these tasks, thereby allowing the large-scale model to execute the tasks in the correct order. In the answer generation and presentation information, the positions of the question and the search content set can be pre-set, and the question, the search content set, and the tasks to be executed can be combined to form a complete input. 【0083】 According to embodiments of this disclosure, by entering the question and search content set into pre-set positions in the answer generation presentation information to form a complete input and setting the order identification of the tasks to be executed, the large-scale model can execute the corresponding text processing tasks based on the determined tasks and order, thereby improving the execution efficiency and controllability of the large-scale model when generating answers. 【0084】 According to embodiments of this disclosure, the task to be performed now may include a content sorting task. Based on the answer generation task to be performed now in the answer generation presentation information, processing the current text corresponding to the search content set and obtaining the processed text includes, based on the content sorting task, performing content processing on the current text and obtaining the processed text with enhanced content. 【0085】 According to embodiments of this disclosure, the content reorganization task may involve performing a series of optimization processes on the current text, including, but not limited to, optimizing the content, order, semantics, etc., of the current text, thereby obtaining a processed text with enhanced content. The processed text with enhanced content is improved in terms of content quality, information content, or readability. 【0086】 According to embodiments of this disclosure, processing the current text with a content organization task can concentrate the processed text on core information, reduce redundant or irrelevant content, and improve readability. 【0087】 According to embodiments of this disclosure, the content sorting task includes a content selection task. Based on the content sorting task, content processing is performed on the current text corresponding to the search content set to obtain the content-enhanced processed text, which includes determining a predetermined number of target subtexts from the current text as the content-enhanced processed text based on the content relevance of each of several subtexts in the current text and the attribute relevance of each of several subtexts in the current text. The content relevance may be determined based on the similarity between the subtext and the question. The attribute relevance may be determined based on subtext tracking information. The subtext tracking information is determined based on the search content set, and the predetermined number is determined based on the content selection task. 【0088】 According to embodiments of this disclosure, multiple subtexts may be generated from results retrieved from different data sources. Multiple subtexts may also be generated from different results retrieved from the same data source. Each subtext may correspond to one search result. For example, multiple subtexts may correspond one-to-one to multiple search contents in a set of search contents. 【0089】 According to embodiments of this disclosure, content relevance can be implicitly associated with a question and multiple subtexts using large-scale model internal vectors. This sorts the subtexts based on the content relevance results, and the top multiple subtexts are designated as a predetermined number of target subtexts. The similarity calculation method used by the large-scale model is not limited herein and is not limited to any method used to evaluate the content relevance between the question and the subtexts. 【0090】 According to embodiments of this disclosure, the tracking information for subtext represents information related to the source of the subtext. This includes, but is not limited to, information such as the distribution site, distributor, distribution time, and the authority and relevance of the distribution site or distributor of the subtext content. Based on the tracking information, subtext with questionable sources can be removed. Subtext that does not meet the deadline for the current question request may also be removed. This results in the selected subtext having greater authority and truthfulness. 【0091】 For example, the current question is "Climate change trends in a certain region in recent years," and one subtext included in the retrieved search content set is from an unverified personal blog, whose source is questionable as it consistently publishes unverified messages. By choosing to remove this subtext when performing the content screening task, the inclusion of inaccurate or misleading information can be avoided. Another included subtext originates from a reliable news organization, but its publication date is 10 years ago, clearly not meeting the time constraint of "recent years." Therefore, by choosing to remove this subtext when performing the content screening task, the current question's deadline requirement can be met. 【0092】 Subtext tracking information can be scored according to one or more of the following rules, such as reliability rules, effectiveness rules, and authority rules, to obtain attribute relevance. However, it is not limited to this. Attribute matching recognition can also be performed on subtext tracking information using large-scale models to obtain attribute relevance. 【0093】 The target relevance can be obtained by weighting attribute relevance and content relevance. Based on the target relevance, a predetermined number of target subtexts can be determined from multiple subtexts. 【0094】 According to the embodiments of this disclosure, a content selection task in a content sorting task is performed on the current text based on the content relevance and attribute relevance of multiple subtexts, ensuring that the resulting processed text is highly relevant to the question and possesses greater truthfulness and authority. 【0095】 According to embodiments of this disclosure, a content sorting task includes a content extraction task. Based on the content sorting task, performing content processing on the current text to obtain processed text with enhanced content includes, based on the content extraction task, performing noise reduction processing on the current text to obtain text with reduced noise. Content extraction is performed on the text with reduced noise to obtain multiple text segments with enhanced hierarchical relationships, which are then used as processed text with enhanced content. 【0096】 According to embodiments of this disclosure, noise reduction processing on the current text may include cleaning the current text and removing unwanted content such as spaces, line breaks, comments, web addresses, and tags. It may also avoid processing the same information twice by recognizing and deleting duplicate text lines or paragraphs. Unused words, duplicate words, and irrelevant words may be removed, and ambiguous words may be removed or replaced. Grammatical errors and spelling mistakes may be corrected. Formatting updates may be performed on the text data so that the text meets specific formatting requirements. Noise reduction processing on the current text can improve the quality and readability of the processed text with enhanced content. 【0097】 According to embodiments of this disclosure, content extraction from noise-reduced text includes extracting entities, keywords, events, and relationships from the text content. 【0098】 Specifically, a large-scale model is pre-trained by training it with data in a database, enabling the model to recognize entity information such as names of people, places, organizations, times, and dates from the current text. Events in the current text, such as meetings, transactions, and natural disasters, are recognized, and key information such as event type, time, location, and participants is extracted. Relationships between entities in the text, such as "A is a subsidiary of B" or "C collaborates with D," are recognized, and the relationship type and participating entities are extracted. The extracted key information, such as entities, keywords, events, and relationships, is organized and output according to a predetermined structure. 【0099】 According to embodiments of this disclosure, multiple target subtexts that have undergone a content selection task can be used as the current text. Multiple target subtexts can be combined to obtain the current text. Each target subtext in the current text is one or more text segments of the current text, but is not limited thereto. A content extraction task may be performed directly without going through a content selection task. Multiple search content in a search content set can be combined to obtain the current text. Each search content is one or more text segments of the current text. 【0100】 According to the embodiments of this disclosure, there may be the same knowledge points among multiple text segments of the current text, as well as knowledge points of different dimensions. By performing noise reduction processing on the current text, duplicate and redundant information can be removed, while different information relevant to the question can be retained. When extracting text segment content, relevant information at different dimensions or granularities of the question can be obtained from different text segments, resulting in higher quality text content after extraction. 【0101】 According to the embodiments of this disclosure, by first performing noise reduction processing on the current text, irrelevant information and noise in the current text can be effectively removed, allowing subsequent content extraction to focus more on key information and improving the accuracy and efficiency of content extraction. 【0102】 According to one embodiment of the present disclosure, content extraction is performed on the noise-reduced text to obtain multiple text segments with enhanced hierarchical relationships, and the resulting processed text with enhanced content may include rearranging the multiple text segments based on the contextual relationships between them in the noise-reduced text, generating identification information to identify the contextual relationships of each of the multiple text segments, and obtaining the processed text with enhanced paragraph hierarchical relationships; and rearranging the multiple phrases based on the contextual relationships between them obtained by dividing the text segments in the processed text with enhanced paragraph hierarchical relationships, generating identification information to identify the contextual relationships of each of the multiple phrases, and obtaining the processed text with enhanced hierarchical relationships. 【0103】 According to embodiments of the present disclosure, after noise reduction processing is performed on the current text and duplicates are removed, the processed text may still have multiple text segments, and if the number of processed text segments is greater than a predetermined text segment threshold, for example, if the number of text segments is greater than 2, the entire text segment can be sorted first, and then the sentences within the text segments can be sorted. 【0104】 According to embodiments of this disclosure, the contextual relationships between multiple text segments include logical order, causal relationships, and summation relationships between the text segments. Based on the logical order, causal relationships, and summation relationships between the text segments, identification information, such as paragraph numbers, can be generated to identify the contextual relationships of each of the multiple text segments. By rearranging the multiple text segments based on the paragraph numbers, their order can be aligned with logic and reading habits. 【0105】 According to embodiments of this disclosure, text segments in processed text with enhanced paragraph hierarchy are divided into multiple phrases for finer-grained analysis. Contextual relationships between multiple phrases include logical order, causal relationships, and transformation relationships. Based on the logical order, causal relationships, and transformation relationships between multiple phrases, identification information, such as phrase numbers, is generated to identify the contextual relationships of each of the multiple phrases. The multiple phrases are then rearranged based on the phrase numbers to align their order with logic and expressive conventions. 【0106】 According to other embodiments of the present disclosure, performing content extraction on noise-reduced text to obtain multiple text segments with enhanced hierarchical relationships as processed text with enhanced content may further include rearranging multiple phrases based on the contextual relationships between the multiple phrases obtained by dividing the noise-reduced text, generating identification information to identify the contextual relationships of each of the multiple phrases, and obtaining processed text with enhanced hierarchical relationships. 【0107】 According to embodiments of this disclosure, after noise reduction processing is performed on the current text, if the number of processed text segments is less than a predetermined text segment threshold, for example, if the number of text segments is 2 or less, all text segments can be divided into multiple phrases. Based on the logical order, causal relationships, and transformation relationships between the multiple phrases, the contextual relationships of the multiple phrases are determined, and a phrase number is generated to identify the contextual relationship of each of the multiple phrases. The multiple phrases are rearranged based on the phrase numbers to make their order more logical and conform to expressive conventions. 【0108】 According to the embodiments of this disclosure, text content is rearranged based on contextual relationships between text segments and phrases, while preserving the original text information, so that the processed text has a clearer and more logical structure. 【0109】 According to the embodiments of this disclosure, a summary generation process is performed on the search content set based on the summary generation task in the answer generation presentation information, and the summary set is obtained as the current text. 【0110】 According to embodiments of this disclosure, each search content in a search content set can be processed by a summarization technology, such as a deep learning-based summarization technology. A concise summary is generated by analyzing the main content, key information, and structure of each search content and removing redundant and secondary information. After the summarization process, each search content corresponds to a single, for example, text segment, and these summaries are combined to form a summary set. The summary set is a summary of the original search content set and contains the information most relevant to the question. 【0111】 For example, a set of summaries can be combined to form the current text. Multiple subtexts in the current text correspond to multiple summaries in the summaries set. A content selection task and content execution task are then sequentially executed on the current text. 【0112】 According to the embodiments of this disclosure, by performing a summary generation process on a search content set, a highly summarized set of summaries is obtained as the current text, significantly simplifying the overall amount of information in the current text and improving the efficiency of subsequent information processing. 【0113】 According to embodiments of this disclosure, the current task to be performed includes a structuring task. Based on the current task to be performed, processing the current text corresponding to the search content set to obtain the processed text includes structuring the current text based on the structuring task to obtain a processed text with enhanced structure. 【0114】 According to embodiments of this disclosure, the structuring task converts unstructured text content into text content in an information format having a clear structure and organization. 【0115】 According to embodiments of this disclosure, structuring the current text includes determining a basic structural framework for the current text based on the text content. Based on the basic structural framework for the current text, the structural elements to be used, such as titles, subtitles, lists, etc., are determined. The content of the current text is redistributed within the structural framework. Information can be organized by titles and subtitles, ensuring that the organized information is clear and easy to understand. Similar or related information points may be organized by lists. 【0116】 According to embodiments of this disclosure, structuring the current text may further include inspecting the structured text, removing redundant and repetitive information, ensuring the accuracy and completeness of the information, and ensuring consistent and logical use of structured elements. The readability of the text can also be enhanced by applying appropriate formatting and typesetting. For example, clear fonts and font sizes can be used to ensure the text is easy to read. The structure of the text may also be enhanced by using appropriate indentation, alignment, identifiers, spaces, etc. 【0117】 According to the embodiments of this disclosure, by applying structuring processing to the current text, the text content can be made clearer and easier to understand, and its readability can be significantly improved. 【0118】 According to embodiments of this disclosure, performing a structuring process on the current text based on a structure tidying task to obtain a processed text with enhanced structure includes performing structure recognition on the current text and obtaining structure recognition information for each of the multiple text segments in the current text. Based on the structure format in the structure tidying task and the structure recognition information for each of the multiple text segments, the format is updated on the current text to obtain a processed text with enhanced structure. 【0119】 According to embodiments of this disclosure, performing structured recognition on the current text includes identifying and recognizing the location of elements such as titles, subtitles, lists, and tables within the current text. The structured recognition information describes the location and role of text segments within the text structure, for example, whether or not they are titles, or whether or not they belong to a subtitle or list. 【0120】 According to embodiments of this disclosure, formatting the current text involves applying a target structural format to the corresponding text segment based on structural recognition information for each text segment. For example, applying a specific font and size to a title, and applying a specific indentation and numbering style to a list. Then, optimization and consistency checks are performed on the formatted text. This ensures that the formatting of all text segments conforms to the requirements of the target structural format and that the overall text style is consistent. 【0121】 According to the embodiments of this disclosure, by updating a predetermined format based on the structural format in the structure recognition and structure organization tasks, it is possible to improve information extraction efficiency and ensure consistency in text structure. 【0122】 Figure 6 schematically shows a template diagram of answer generation and presentation information according to an embodiment of the present disclosure. 【0123】 As shown in Figure 6, the answer generation and presentation information includes several parts, such as the task role 610, the task description 620, the task example 630, the designated location 640 in the search content set, and the designated location 650 for the question. The task role 610 is used to define the role of the large model as a question answering expert. By performing the task in the task description 620, the large model can process the current text obtained by searching based on the question entered in the designated location 650 and output an answer that satisfies the specified conditions based on the example 630. 【0124】 Specifically, the large-scale model first performs a summary generation process on the search content set obtained by searching based on the question, and obtains the summary set as the current text. It then performs sorting and filtering on the current text to obtain sorted text containing multiple target subtexts, and performs noise reduction processing on the sorted text to obtain noise-reduced text containing multiple text segments. It then performs content organization on the multiple text segments in the noise-reduced text to obtain content-enhanced text. Finally, it performs structural organization on the content-enhanced text based on the output format requirements and outputs the final answer. Note that the above order is illustrative, and appropriate deletions or additions of new processing tasks may be made to the above tasks, and this disclosure is not limited thereto. 【0125】 Figure 7 schematically shows the effect of the processed text according to the embodiment of this disclosure. As shown in Figure 7, the processed text obtained based on the answer generation method of this disclosure has a title, subtitle and related information, is rich in content, has many information points, has a clear structure and hierarchy. 【0126】 According to the embodiments of this disclosure, the processed text is evaluated based on the evaluation information in the presented information, and an evaluation result is obtained to indicate whether or not the processed text satisfies predetermined conditions. 【0127】 According to the embodiments of this disclosure, the evaluation information in the answer generation and presentation information may be a set of predefined reference information, such as a reference text sample, which represents a text format and content that meets specific requirements or criteria. However, it is not limited to this. It may also be a set of predefined evaluation metrics, such as an evaluation metric that the processed text meets the format, an evaluation metric that meets the richness of the content, or an evaluation metric that meets the depth of the content. The evaluation metrics and reference information may be combined to form the evaluation information. Any evaluation criterion that can be used to provide feedback to the user as an answer to a question is sufficient, as long as it can evaluate whether the processed text meets expectations. 【0128】 The specified conditions may correspond to evaluation information. If the evaluation information includes reference information, the specified conditions may include a goodness-of-fit threshold. The processed text and the reference information can be matched to determine the degree of fit between them. If the degree of fit is greater than the goodness-of-fit threshold, an evaluation result that satisfies the specified conditions is obtained. If the degree of fit is less than or equal to the goodness-of-fit threshold, an evaluation result that does not satisfy the specified conditions is obtained. 【0129】 If the evaluation information includes evaluation indicators, the specified conditions may also include evaluation values. The processed text can be evaluated using a predetermined set of multiple evaluation indicators, sub-evaluation values ​​corresponding to the multiple evaluation indicators can be obtained, and the evaluation value can be obtained by weighting and adding the multiple sub-evaluation values. If the evaluation value is greater than the evaluation value threshold, an evaluation result that satisfies the specified conditions is obtained. If the evaluation value is less than or equal to the evaluation value threshold, an evaluation result that does not satisfy the specified conditions is obtained. 【0130】 The system can output processed text as an answer that simultaneously satisfies both the evaluation threshold and the goodness-of-fit threshold. If one of these thresholds is not met, the system can repeatedly execute content organization tasks and structure organization tasks based on the question, answer generation information, and search content set to obtain an answer that matches the expectations. 【0131】 According to the embodiments of this disclosure, by using evaluation information as evaluation criteria, it is possible to ensure that a large-scale model generates processed text that conforms to predetermined conditions, thereby enabling controllable guidance based on the output results using the evaluation information. 【0132】 Selectively, evaluation indicators in the evaluation information can be set for each task to be executed, the completion status of each task after execution can be evaluated, and it can be determined whether or not the predetermined conditions are met. If the predetermined conditions are not met, the task can be repeated. Reference information in the evaluation information can be set as the last task to be executed, and the processed text can be evaluated using the final version of the reference information to ensure the completeness and consistency of the final output answer. 【0133】 According to embodiments of this disclosure, by adding sequence identification to instruct the execution order of tasks in the process of executing a question answering method based on a large-scale model, the execution order of tasks in the large-scale model can be controlled as it processes multiple tasks. Furthermore, by adding evaluation information to the answer generation and presentation information, the large-scale model can accurately evaluate the results after task execution and avoid generating hallucinatory answers. 【0134】 Embodiments of this disclosure further provide an agent configured to perform a question-answering method based on a large-scale model, as shown in Figure 2. 【0135】 The agent is a high-level artificial intelligence system that uses a large-scale model as its core inference engine. It not only possesses the language understanding and generation capabilities of a large-scale model, but also efficiently and flexibly solves various complex questions, further unlocking the machine intelligence contained within the large-scale language model to provide users with more accurate and personalized services. 【0136】 Figure 8 schematically shows an agent according to an embodiment of the present disclosure. As shown in Figure 8, the agent 800 may include an input unit 810, a control unit 820, a storage unit 830, an arithmetic unit 840, and an output unit 850. 【0137】 The input unit 810 receives or senses queries, requests, commands, questions, signals, or data from external sources, such as a user or the external environment, and converts them into formatted information that the agent can understand and process. The input unit 810 is the first part of the interaction between the agent 800 and the outside world, enabling the agent 800 to efficiently and accurately acquire and respond to the necessary "sensory" information from the outside. 【0138】 In the example, the input unit 810 can perform an operation to obtain a question related to the method shown in Figure 2. 【0139】 The control unit 820 is the core support for agent 800 to handle complex task capabilities. The control capabilities of the control unit 820 include four aspects: planning capability, action capability, evaluation capability, and reconsideration capability. In the example, the control unit can determine the task to be performed using its planning capability. Using its action capability, it can execute operations S220 and S230 shown in Figure 2, process the current text, and obtain the processed text. Using its evaluation capability, it evaluates the processed document output based on the evaluation information and determines whether the processed text can be used as an answer to a question. Using its reconsideration capability, it re-executes operations S220 and S230 if it is determined that the processed text does not meet predetermined conditions. In the example, the control unit 820 continuously interacts with the storage unit 830, the arithmetic unit 840, and / or the output unit 850 during operation. However, in the embodiments of this disclosure, the control unit 820 initiates communication with the storage unit 830, the arithmetic unit 840, and / or the output unit 850 as a single initiator, and there is no communication coupling between the storage unit 830, the arithmetic unit 840, and the output unit 850. 【0140】 In the example, the performance of the control unit 820 may be related to the large-scale model on which the agent 800 is based. To fully utilize the capabilities of the large-scale model, the internal structure of the control unit 820 may be designed to be highly configurable and expandable to accommodate various different types of tasks and needs in real-world scenarios. 【0141】 The memory unit 830 can store information such as history interactions and event flows. This stored information may also be stored in the memory unit 830. 【0142】 In this example, after acquiring input information such as a question, agent 800 can determine whether or not it is necessary to trigger a search operation. If agent 800 determines that it is not necessary to trigger a search operation, it may acquire the answer corresponding to the question and feed it back from the storage unit 830 to the control unit 820. The control unit 820 can then use the fed-back answer to transmit it to the output unit 850. 【0143】 The arithmetic unit 840 can be considered a predefined tool library. Plugin tools, function tools, interface tools, and model tools as described above may be included in the arithmetic unit 840. 【0144】 In the example, if agent 800 determines that the execution information includes tool information, it may retrieve that tool information from the arithmetic unit 840 and feed it back to the control unit 820. The control unit 820 can use the fed-back tool information, for example a search engine, to search for multiple rewrite questions, obtain a set of search content, generate presentation information based on the set of search content, questions, and answers obtained from the memory unit 830, perform the operations shown in Figure 2, and obtain the answers. It then transmits the answers to the arithmetic unit 840 so that the arithmetic unit executes the output unit 850. While large-scale models have excellent language understanding and generation capabilities, it is understood that, like humans, the tasks they can solve without using any tools are very limited. When agent 800 is given the ability to call tools, tasks such as mathematical calculations by a calculator, data analysis by data analysis software, and weather forecasting by a search engine can be realized. 【0145】 The agent 800 according to the embodiments of this disclosure can easily and effectively improve intelligence, flexibility, and versatility. 【0146】 Figure 9 schematically shows a block diagram of a question answering device according to an embodiment of the present disclosure. As shown in Figure 9, the question answering device 900 includes an input module 910, an input module 920, and a processing module 930. 【0147】 The input module 910 inputs the question, the search content set, and the answer generation presentation information into the large-scale model in response to the search content set retrieved based on the question. 【0148】 The processing module 920 processes the current text corresponding to the search content set based on the current task to be performed and the question in the answer generation presentation information, retrieves the processed text, and determines the current task to be performed based on the task execution order in the answer generation presentation information. 【0149】 The first generation module 930 obtains the answer to the question based on the processed text if it determines that the processed text satisfies predetermined conditions. 【0150】 According to the embodiments of this disclosure, the processing module includes a content processing submodule. The content processing submodule performs content processing on the current text based on the content cleanup task and obtains the processed text with enhanced content. 【0151】 According to the embodiments of this disclosure, the content processing submodule includes a content sorting unit. 【0152】 The content selection unit determines a predetermined number of target text segments from the current text as content-enhanced processed text, based on the content relevance of each of the multiple text segments in the current text and the attribute relevance of each of the multiple text segments in the current text. The content relevance is determined based on the similarity between the text segment and the question, the attribute relevance is determined based on the tracking information of the text segment, the tracking information of the text segment is determined based on the search content set, and the predetermined number is determined based on the content selection task. 【0153】 According to embodiments of this disclosure, the content processing submodule further includes a noise reduction unit and an extraction unit. 【0154】 The noise reduction unit performs noise reduction processing on the current text based on the content extraction task and obtains the noise-reduced text. 【0155】 The extraction unit performs content extraction on the noise-reduced text, obtaining multiple text segments with enhanced hierarchical relationships as processed text with enhanced content. 【0156】 According to embodiments of the present disclosure, the extraction unit includes a text segment processing subunit and a first phrase processing subunit. 【0157】 The text segment processing subunit rearranges multiple text segments based on the contextual relationships between them in the noise-reduced text, generates identification information to identify the contextual relationships of each of the multiple text segments, and obtains processed text with enhanced paragraph hierarchy relationships. 【0158】 The first word processing subunit rearranges multiple words based on the contextual relationships between them obtained by dividing text segments in the processed text with enhanced paragraph hierarchy, generates identification information to identify the contextual relationships of each of the multiple words, and obtains the processed text with enhanced hierarchy. 【0159】 According to embodiments of the present disclosure, the extraction unit further includes a second phrase processing subunit. 【0160】 The second word processing subunit rearranges the multiple words obtained by splitting the noise-reduced text based on the contextual relationships between them, generates identification information to identify the contextual relationships of each of the multiple words, and obtains a processed text with enhanced hierarchical relationships. 【0161】 According to embodiments of the present disclosure, the question answering device 900 further includes a summarization processing module. The summarization processing module performs a summarization process on the search content set based on the summarization task in the presented information and retrieves the summarization set as the current text. 【0162】 According to embodiments of this disclosure, the processing module 920 further includes a structural processing submodule. 【0163】 The structuring submodule performs structuring on the current text based on the structuring task, and obtains the processed text with enhanced structure. 【0164】 According to embodiments of this disclosure, the structure processing submodule includes a structure recognition unit and a structure update unit. 【0165】 The structure recognition unit performs structure recognition on the current text and obtains structure recognition information for each of the multiple text segments in the current text. 【0166】 The structure update unit updates the format of the current text based on the structure format and structure recognition information of each of the multiple text segments in the structure organization task, and obtains processed text with an enhanced structure. 【0167】 According to embodiments of this disclosure, the question answering device 900 further includes an evaluation module. The evaluation module evaluates the processed text based on the evaluation information in the presented information and obtains an evaluation result that indicates whether the processed text meets predetermined conditions. 【0168】 According to embodiments of this disclosure, the evaluation information may include at least one of an evaluation index and reference information. 【0169】 According to embodiments of this disclosure, the input module 910 further includes a first input submodule. 【0170】 The first input submodule inputs the question and search content sets into pre-configured locations in the answer generation presentation information, and then inputs them into the large-scale model. Here, the presentation information includes multiple tasks, and each task is assigned an order identifier to indicate the order in which the tasks should be executed. 【0171】 According to embodiments of the present disclosure, the question answering device further includes an analysis module and a question rewriting module. 【0172】 The analysis module performs search trigger analysis on a question based on the search trigger recognition task in the question analysis presentation information, and obtains the search trigger analysis results. 【0173】 The question rewriting module, when the search trigger analysis results indicate that a search operation should be triggered, rewrites the question based on the question rewriting task in the question analysis presentation information to obtain multiple rewritten questions and retrieve the search content set based on the multiple rewritten questions. 【0174】 According to embodiments of the present disclosure, the question answering device 900 further includes a second generation module. 【0175】 The second generation module generates answers to questions based on the question if the search trigger analysis results indicate that a search operation does not need to be triggered. 【0176】 According to the embodiments of this disclosure, the question rewriting module includes a question rewriting submodule. 【0177】 The Question Rewrite submodule rewrites a question based on rewrite rules that match the question type of the question in the Question Rewrite task, thereby obtaining multiple rewritten questions. 【0178】 According to embodiments of the present disclosure, the question rewriting module further comprises a first rewriting unit and a second rewriting unit. 【0179】 The first rewrite unit rewrites questions based on the question rewrite task and obtains multiple initial rewritten questions. 【0180】 The second rewriting unit obtains multiple rewrite questions based on multiple initial rewrite questions if the relevance between each of the multiple initial rewrite questions and the question satisfies the relevance threshold in the question rewriting task. 【0181】 According to embodiments of the present disclosure, the input module further includes a second input submodule. 【0182】 The second input submodule inputs the questions into the large-scale model by placing them in pre-set locations in the question analysis presentation information, where the search trigger recognition task and the question rewriting task in the question analysis presentation information each have sequence identifiers added to indicate the order in which the tasks should be executed. 【0183】 According to embodiments of the present disclosure, the present disclosure further provides electronic devices, readable storage media, and computer programs. 【0184】 According to embodiments of the present disclosure, the electronic device includes at least one processor and a memory that is communicated with the at least one processor, the memory storing instructions that can be executed by the at least one processor, the instructions being executed by the at least one processor, and enabling the at least one processor to perform the above-described method. 【0185】 According to embodiments of the present disclosure, a non-temporary computer-readable storage medium storing computer instructions is used to cause a computer to perform the above-described method. 【0186】 According to embodiments of this disclosure, a computer program, when executed by a processor, achieves the above method. 【0187】 Figure 10 shows an exemplary block diagram for carrying out an electronic device 1000 of an embodiment of the present disclosure. The electronic device is intended to display various forms of digital computers, such as laptop computers, desktop computers, workbenches, personal digital assistants, servers, blade servers, mainframe computers, and other appropriate computers. The electronic device may further display various forms of mobile devices, such as personal digital assistants, mobile phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are illustrative and not intended to limit the implementation of the present disclosure as described herein and / or requested. 【0188】 As shown in Figure 10, the device 1000 includes a computing unit 1001, which can perform various appropriate operations and processes based on computer programs stored in read-only memory (ROM) 1002 or computer programs loaded from storage unit 1008 into random access memory (RAM) 1003. The RAM 1003 can also store various programs and data necessary for the operation of the device 1000. The computing unit 1001, ROM 1002, and RAM 1003 are interconnected via a bus 1004. An input / output (I / O) interface 1005 is also connected to the bus 1004. 【0189】 Multiple components in device 1000 are connected to an input / output (I / O) interface 1005 and include, for example, an input unit 1006 such as a keyboard or mouse; an output unit 1007 such as various types of displays or speakers; a storage unit 1008 such as a magnetic disk or optical disk; and a communication unit 1009 such as a network card, modem, or wireless communication transceiver. The communication unit 1009 enables device 1000 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks. 【0190】 The computing unit 1001 may be various general-purpose and / or dedicated processing modules having processing and computing capabilities. Some examples of the computing unit 1001 include, but are not limited to, a central processing unit (CPU), a GPU (Graphics Processing Unit), various dedicated artificial intelligence (AI) computing chips, computing units that execute various machine learning model algorithms, a DSP (Digital Signal Processor), and any suitable processor, controller, microcontroller, etc. The computing unit 1001 performs each of the methods and processes described above, such as a question answering method. For example, in some embodiments, the question answering method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as a memory unit 1008. In some embodiments, part or all of the computer program may be loaded and / or installed into the device 1000 via ROM 1002 and / or communication unit 1009. When the computer program is loaded into RAM 1003 and executed by the computing unit 1001, one or more steps of the question answering method described above may be performed. Alternatively, in another embodiment, the computing unit 1001 may be configured to perform the question answering method in any other suitable form (e.g., via firmware). 【0191】 Various embodiments of the systems and technologies described herein may be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-chip (SOCs), complex-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may be implemented in one or more computer programs, which can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, which may include receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device. 【0192】 Program code for carrying out the methods of this disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a dedicated computer, or other programmable data processing device, so that when the program code is executed by the processor or controller, the functions / operations defined in the flowcharts and / or block diagrams are performed. The program code may be executed entirely on a device, partially on a device, partially on a device as a standalone software package, partially on a remote device, or entirely on a remote device or server. 【0193】 In the context of this disclosure, a machine-readable medium may be a tangible medium that contains or stores programs used in or in combination with an instruction execution system, device, or electronic device. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or electronic devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include one or more wired electrical connections, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing. 【0194】 To provide user interaction, a computer may be made to implement the systems and techniques described herein, the computer comprising a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) and a keyboard and pointing device (e.g., a mouse or trackball), the user may provide input to the computer via the keyboard and pointing device. Other types of devices may further provide user interaction, for example, feedback provided to the user may be any form of sensing feedback (e.g., visual feedback, auditory feedback, or haptic feedback), and input from the user may be received in any form (including voice input, speech input, or haptic input). 【0195】 The systems and technologies described herein can be implemented in computing systems including background components (e.g., a data server), computing systems including middleware components (e.g., an application server), computing systems including front-end components (e.g., a user computer having a graphical user interface or a web browser, through which the user can interact with embodiments of the systems and technologies described herein), or in computing systems including any combination of such background components, middleware components, or front-end components. Components of the system can be connected to one another by digital data communication (e.g., a communication network) in any form or medium. Examples of communication networks include, but are not limited to, local area networks (LANs), wide area networks (WANs), and the Internet. 【0196】 A computer system may include clients and servers. Clients and servers are generally geographically separated and typically interact via a communication network. The client-server relationship is generated by a computer program running on the relevant computer that has a client-server relationship. The server may be a cloud server, a server in a distributed system, or a server combined with a blockchain. 【0197】 It should be understood that various forms of flows shown above may be used, and steps may be rearranged, added, or deleted. For example, each step described in the present invention may be performed in parallel, sequentially, or in a different order, as long as the desired results of the proposed invention of this disclosure can be achieved. 【0198】 The specific embodiments described above do not limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, subcombinations, and substitutions are possible depending on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

[Claim 1] A computer-based question answering method, which is based on a large-scale model, In response to the search content set retrieved based on the question, the question, the search content set, and the answer generation and presentation information are input into a large-scale model. In the aforementioned large-scale model, Based on the current task to be performed in the answer generation presentation information and the question, the current text corresponding to the search content set is processed, and the processed text is obtained, and the current task to be performed is an operation determined based on the task execution order in the answer generation presentation information, If it is determined that the processed text satisfies the predetermined conditions, the system will perform the operation of obtaining the answer to the question based on the processed text. A question-answering method based on large-scale models. [Claim 2] The aforementioned tasks to be performed now include content organization tasks, Based on the answer generation task to be performed in the answer generation presentation information, processing the current text corresponding to the search content set and obtaining the processed text is: This includes performing content processing on the current text based on the content sorting task and obtaining the processed text with enhanced content. The method according to claim 1. [Claim 3] The aforementioned content organization task includes a content selection task, Based on the aforementioned content organization task, content processing is performed on the current text corresponding to the search content set, and the processed text with enhanced content is obtained. Based on the content relevance of each of the multiple subtexts in the current text and the attribute relevance of each of the multiple subtexts in the current text, a predetermined number of target subtexts are determined from the current text as the content-enhanced processed text, wherein the content relevance is determined based on the similarity between the subtext and the question, the attribute relevance is determined based on the tracking information of the subtext, the tracking information of the subtext is determined based on the search content set, and the predetermined number is determined based on the content selection task. The method according to claim 2. [Claim 4] The aforementioned content organization task includes a content extraction task, Based on the aforementioned content organization task, performing content processing on the current text and obtaining the processed text with enhanced content is: Based on the content extraction task, noise reduction processing is performed on the current text, and the text after noise reduction is obtained. This includes performing content extraction on the noise-reduced text to obtain multiple text segments with enhanced hierarchical relationships, and using this as the processed text with enhanced content. The method according to claim 2 or 3. [Claim 5] Performing content extraction on the noise-reduced text to obtain multiple text segments with enhanced hierarchical relationships, and then using this as the processed text with enhanced content, is: Based on the contextual relationships between multiple text segments in the noise-reduced text, the multiple text segments are rearranged to generate identification information for identifying the contextual relationships of each of the multiple text segments, and a processed text with enhanced paragraph hierarchy relationships is obtained. The process includes: rearranging multiple phrases based on contextual relationships between them to generate identification information for identifying the contextual relationships of each of the multiple phrases; obtaining a processed text with enhanced hierarchical relationships; and obtaining the multiple phrases by dividing the text segments in the processed text with enhanced paragraph hierarchical relationships. The method according to claim 4. [Claim 6] Performing content extraction on the noise-reduced text to obtain multiple text segments with enhanced hierarchical relationships, and then using this as the processed text with enhanced content, is: This includes rearranging the multiple phrases obtained by dividing the noise-reduced text based on the contextual relationships between them, generating identification information to identify the contextual relationships of each of the multiple phrases, and obtaining a processed text with enhanced hierarchical relationships. The method according to claim 4. [Claim 7] This further includes performing a summary generation process on the search content set based on the summary generation task in the answer generation presentation information, and obtaining the summary set as the current text. The method according to any one of claims 1 to 3. [Claim 8] The tasks to be performed at present include structural organization tasks, Based on the current task to be performed, processing the current text corresponding to the search content set and obtaining the processed text is: This includes performing a structuring process on the current text based on the aforementioned structuring task, and obtaining a processed text with an enhanced structure. The method according to any one of claims 1 to 3. [Claim 9] Based on the aforementioned structural organization task, performing a structuring process on the current text and obtaining the processed text with an enhanced structure is: Performing structured recognition on the current text and obtaining structured recognition information for each of the multiple text segments in the current text, This includes updating the format of the current text based on the structural format in the structural organization task and the structural recognition information of each of the multiple text segments, and obtaining the processed text with an enhanced structure. The method according to claim 8. [Claim 10] The process further includes evaluating the processed text based on the evaluation information in the answer generation and presentation information, and obtaining an evaluation result to indicate whether the processed text satisfies the predetermined conditions. The method according to any one of claims 1 to 3. [Claim 11] The aforementioned evaluation information includes at least one of the following: an evaluation index and reference information. The method according to claim 10. [Claim 12] Inputting the aforementioned questions, the search content set, and the answer generation and presentation information into a large-scale model is, The aforementioned questions and the search content set are entered into pre-set positions in the answer generation and presentation information, respectively, and input into the large-scale model. The answer generation and presentation information includes multiple tasks, and the tasks are assigned sequence identifiers to indicate the execution order of the tasks. The method according to any one of claims 1 to 3. [Claim 13] In response to the received question, the question and the question analysis presentation information are input into the large-scale model. In the aforementioned large-scale model, Based on the search trigger recognition task in the aforementioned question analysis presentation information, the operation involves performing a search trigger analysis on the question and obtaining the search trigger analysis results. If the search trigger analysis results indicate that a search operation needs to be triggered, the operation further includes: rewriting the question based on the question rewriting task in the question analysis presentation information, obtaining a set of rewritten questions, and retrieving the search content set based on the set of rewritten questions. The method according to any one of claims 1 to 3. [Claim 14] If the search trigger analysis results indicate that there is no need to trigger a search operation, the process further includes generating an answer to the question based on the question. The method according to claim 13. [Claim 15] Based on the question rewriting task in the question analysis presentation information, rewriting the question and obtaining multiple rewritten questions is: The question rewriting task includes rewriting the question based on a rewriting rule that matches the question type of the question, thereby obtaining the plurality of rewritten questions. The method according to claim 13. [Claim 16] Based on the question rewriting task in the question analysis presentation information, rewriting the question and obtaining multiple rewritten questions is: Based on the aforementioned question rewriting task, the question is rewritten to obtain multiple initial rewritten questions, If the relationship between each of the plurality of initial rewrite questions and the question satisfies the relationship threshold in the question rewriting task, then the plurality of rewrite questions are obtained based on the plurality of initial rewrite questions. The method according to claim 15. [Claim 17] Inputting the aforementioned questions and question analysis presentation information into the large-scale model means that This includes entering the aforementioned questions into pre-set locations in the question analysis presentation information, inputting them into the large-scale model, and adding sequence identifiers to the search trigger recognition task and the question rewriting task in the question analysis presentation information to indicate the execution order of the tasks. The method according to claim 16. [Claim 18] A question answering device based on a large-scale model, An input module that inputs the question, the search content set, and the answer generation and presentation information into a large-scale model in response to the search content set retrieved based on the question, Based on the current task to be performed in the answer generation presentation information and the question, the processing module processes the current text corresponding to the search content set, obtains the processed text, and the current task to be performed is determined based on the task execution order in the answer generation presentation information. If it is determined that the processed text satisfies predetermined conditions, a first generating module that obtains the answer to the question based on the processed text includes: A question answering system based on a large-scale model. [Claim 19] The aforementioned tasks to be performed now include content organization tasks, The aforementioned processing module is The content processing submodule includes a content processing submodule that performs content processing on the current text based on the content sorting task and obtains processed text with enhanced content. The apparatus according to claim 18. [Claim 20] The aforementioned content organization task includes a content selection task, The aforementioned content processing submodule is: Based on the content relevance of each of the multiple subtexts in the current text and the attribute relevance of each of the multiple subtexts in the current text, a predetermined number of target subtexts are determined from the current text as the content-enhanced processed text, wherein the content relevance is determined based on the similarity between the subtext and the question, the attribute relevance is determined based on the tracking information of the subtext, the tracking information of the subtext is determined based on the search content set, and the predetermined number includes content selection units determined based on the content selection task. The apparatus according to claim 19. [Claim 21] The aforementioned content organization task includes a content extraction task, The aforementioned content processing submodule is: A noise reduction unit performs noise reduction processing on the current text based on the content extraction task and obtains the text after noise reduction. The extraction unit further includes a unit that performs content extraction on the noise-reduced text and obtains multiple text segments with enhanced hierarchical relationships as the processed text with enhanced content. The apparatus according to claim 19 or 20. [Claim 22] The extraction unit is A text segment processing subunit that rearranges the multiple text segments based on the contextual relationships between them in the noise-reduced text, generates identification information to identify the contextual relationships of each of the multiple text segments, and obtains processed text with enhanced paragraph hierarchy relationships. The process includes: rearranging multiple phrases based on contextual relationships between them to generate identification information for identifying the contextual relationships of each of the multiple phrases; obtaining a processed text with enhanced hierarchical relationships; and the multiple phrases comprising a first phrase processing subunit obtained by dividing the text segments in the processed text with enhanced paragraph hierarchical relationships. The apparatus according to claim 21. [Claim 23] The extraction unit is The second phrase processing subunit further includes a second phrase processing subunit that, based on the contextual relationships between multiple phrases obtained by dividing the noise-reduced text, rearranges the multiple phrases to generate identification information for identifying the contextual relationships of each of the multiple phrases, and obtains a processed text with enhanced hierarchical relationships. The apparatus according to claim 21. [Claim 24] The summarization processing module further includes a module that performs a summary generation process on the search content set based on the summary generation task in the answer generation presentation information and obtains the summary set as the current text. The apparatus according to any one of claims 18 to 20. [Claim 25] The tasks to be performed at present include structural organization tasks, The aforementioned processing module is The structure processing submodule further includes a structure processing submodule that performs structure processing on the current text based on the structure processing task and obtains a processed text with enhanced structure. The apparatus according to any one of claims 18 to 20. [Claim 26] The aforementioned structural processing submodule is A structure recognition unit performs structure recognition on the current text and obtains structure recognition information for each of the multiple text segments in the current text. Includes a structure update unit that updates the format of the current text based on the structure format in the structure organization task and the structure recognition information of each of the multiple text segments, and obtains the processed text with an enhanced structure. The apparatus according to claim 25. [Claim 27] The evaluation module further includes an evaluation module that evaluates the processed text based on the evaluation information in the answer generation and presentation information and obtains an evaluation result indicating whether the processed text satisfies the predetermined conditions. The apparatus according to any one of claims 18 to 20. [Claim 28] The aforementioned evaluation information includes at least one of the following: an evaluation index and reference information. The apparatus according to claim 27. [Claim 29] The aforementioned input module is The aforementioned questions and the search content set are entered into pre-set positions in the answer generation and presentation information, respectively, and input into the large-scale model. The answer generation and presentation information further includes a first input submodule which includes multiple tasks, and each task is endowed with an order identifier to indicate the task execution order. The apparatus according to any one of claims 18 to 20. [Claim 30] A question answering device based on a large-scale model, In response to the received question, the question and the question analysis presentation information are input into the large-scale model. An analysis module that performs a search trigger recognition task on the question based on the question analysis presentation information and obtains the search trigger analysis results, If the search trigger analysis results indicate that a search operation needs to be triggered, the question rewrite module further includes: rewriting the question based on the question rewriting task in the question analysis presentation information to obtain multiple rewritten questions, and obtaining the search content set based on the multiple rewritten questions. The apparatus according to any one of claims 18 to 20. [Claim 31] If the search trigger analysis results indicate that there is no need to trigger a search operation, the module further includes a second generation module that generates an answer to the question based on the question. The apparatus according to claim 30. [Claim 32] The aforementioned question rewriting module, The question rewriting submodule includes a question rewriting submodule that rewrites the question based on a rewriting rule that matches the question type of the question in the question rewriting task, thereby obtaining the multiple rewritten questions. The apparatus according to claim 30. [Claim 33] The aforementioned question rewriting module, A first rewrite unit that rewrites the question based on the question rewrite task and obtains a plurality of initial rewrite questions, If the relationship between each of the plurality of initial rewrite questions and the question satisfies the relationship threshold in the question rewrite task, a second rewrite unit that acquires the plurality of rewrite questions based on the plurality of initial rewrite questions further includes: The apparatus according to claim 32. [Claim 34] The aforementioned input module is The aforementioned questions are entered into pre-set positions in the question analysis presentation information and input into the large-scale model, and the search trigger recognition task and the question rewriting task in the question analysis presentation information are further provided with a second input submodule to indicate the task execution order. The apparatus according to claim 33. [Claim 35] It is an electronic device, At least one processor, Includes a memory that is communicably connected to at least one of the processors, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the method according to any one of claims 1 to 3. electronic equipment. [Claim 36] A non-temporary computer-readable storage medium in which computer instructions are stored, The computer instruction is used to cause a computer to perform the method described in any one of claims 1 to 3. A non-temporary, computer-readable storage medium. [Claim 37] A computer program, when executed by a processor, that implements the method described in any one of claims 1 to 3. [Claim 38] He is an agent, The method described in any one of claims 1-3 is configured to perform the following: It includes an input unit, a control unit, a memory unit, an arithmetic unit, and an output unit. agent.

Images